Advancing Lung Cancer Diagnosis: A Hybrid Feature Fusion Approach with Attention Mechanisms and Vision Transformer
Partha Pratim Saha, Mohammed Mahbubur Rahaman, Md Tohfatul Islam, Nahid Imtiaz Chowdhury
Abstract
Lung cancer remains one of the leading causes of cancer-related deaths worldwide, with early detection being crucial for improving survival rates. This study investigates the use of advanced deep learning models, including DenseNet121 with Convolutional Block Attention Module (CBAM) and VGG16 with Squeeze-and-Excitation (SE), for lung cancer classification using CT scan images from the IQ-OTH/NCCD dataset. The dataset comprises images categorized into normal, benign, and malignant classes. By incorporating attention mechanisms such as CBAM and SE, the models were able to focus on key regions of the images, significantly enhancing feature extraction and improving model performance. DenseNet121 with CBAM achieved an accuracy of 97.37% and a loss of 0.0771, while VGG16 with SE achieved an accuracy of 98.48% and a loss of 0.0509. The proposed Unified Ensemble model, which integrates the strengths of both DenseNet121 and VGG16 through feature fusion and attention mechanisms, achieved the highest accuracy of 98.61% with a loss of 0.0445. This model demonstrated superior classification performance and generalization across all categories, outperforming the Vision Transformer (ViT), which achieved an accuracy of 98.07%. Although ViT showed promising results, it requires large amounts of data to achieve optimal performance, highlighting a key limitation of transformer-based models. In contrast, the Unified Ensemble model provides a more efficient solution, delivering high accuracy with fewer data and demonstrating the innovation of combining multiple models with attention mechanisms to improve both performance and generalization in lung cancer diagnosis.